0.31.0
AIs tend not to see the images, which makes the test inconsistent. We addressed this with several approaches:
- Reduced context size by optimizing prompts
- Added image referencing in the UI_STATE prompt
Breaking Change for Users Using userPromptTemplate
To make the project and prompts consistent, we have renamed {{COMMAND_TEMPLATES}} to {{ACTION_TEMPLATES}}. This will generate an error so you can identify if you are using the incorrect syntax.
This is the latest userPromptTemplate template.
<GOAL>{{USER_INPUT_GOAL}}</GOAL>
<STEP>
Current step: {{CURRENT_STEP}}
Max step: {{MAX_STEP}}
<PREVIOUS_STEPS>
{{STEPS}}
</PREVIOUS_STEPS>
</STEP>
<UI_STATE>
Please refer to the image above.
<ELEMENTS>
{{UI_ELEMENTS}}
</ELEMENTS>
<FOCUSED_TREE>
{{FOCUSED_TREE}}
</FOCUSED_TREE>
</UI_STATE>
<INSTRUCTIONS>
Based on the above, decide on the next action to achieve the goal. Please ensure not to repeat the same action. The action must be one of the following:
</INSTRUCTIONS>
<ACTION_OPTIONS>
{{ACTION_TEMPLATES}}
</ACTION_OPTIONS>
What's Changed
- Improve user prompts by adding default messages for empty focused tree and UI elements by @takahirom in #185
- Refactor agent command types to agent action types for consistency in naming by @takahirom in #186
Full Changelog: 0.30.0...0.31.0